PyTorch Price Forecasting

A machine learning system for cryptocurrency price forecasting using LSTM with attention mechanism.

Features

• Fetches historical price data from Pyth Network API
• Processes OHLCV (Open, High, Low, Close, Volume) data
• Calculates technical indicators (RSI, MACD, Bollinger Bands, Moving Averages)
• Uses LSTM with attention mechanism for price prediction
• Predicts next 24 hours of price movements (288 5-minute intervals)
• Saves trained models in organized directory structure
• Real-time forecasting system with live price updates every minute

Setup

1. Install dependencies:

pip install -r requirements.txt

2. Check and fetch training data:

python check_and_fetch_data.py

3. Train the models:

python train_models.py

4. Start real-time forecasting:

python real_time_forecast.py

Data Format

The system expects CSV files with the following columns:

• timestamp: DateTime in format 'YYYY-MM-DD HH:MM:SS'
• open: Opening price
• high: Highest price
• low: Lowest price
• close: Closing price
• volume: Trading volume

API Response Format

The system handles Pyth Network API responses in the format:

{
  "s": "ok",
  "t": [1735707300, 1735707600],
  "o": [93447.18141577, 93357.85438592],
  "h": [93471.38962908, 93390.0],
  "l": [93325.2474621, 93345.19861641],
  "c": [93357.87633587, 93383.75891115],
  "v": [0.0, 0.0]
}

Model Architecture

• Input: 60 time steps of OHLCV data + technical indicators
• LSTM Layers: 2 layers with 128 hidden units
• Attention: Multi-head attention mechanism
• Output: 288 predictions (24 hours of 5-minute intervals)
• Loss: Mean Squared Error (MSE)

Technical Indicators

The model includes the following technical indicators:

• RSI (Relative Strength Index)
• MACD (Moving Average Convergence Divergence)
• Bollinger Bands
• Moving Averages (5, 10, 20, 50 periods)
• Price change ratios
• Volatility measures
• Time-based features

Directory Structure

pytorch-price-forecast/
├── training_data/
│   ├── bitcoin_5min.csv
│   ├── ethereum_5min.csv
│   └── xau_5min.csv
├── models/
│   ├── best_model_bitcoin.pth
│   ├── best_model_ethereum.pth
│   └── best_model_xau.pth
├── forecasts/
│   ├── bitcoin_forecast_20240115_1430.csv
│   ├── ethereum_forecast_20240115_1430.csv
│   └── xau_forecast_20240115_1430.csv
├── fetch_training_data.py
├── model_train.py
├── real_time_forecast.py
├── test_data_fetch.py
├── test_real_time.py
└── requirements.txt

Usage

Training

from model_train import PricePredictor

# Initialize predictor
predictor = PricePredictor(sequence_length=60, prediction_horizon=288)

# Prepare data
file_paths = {
    'bitcoin': './training_data/bitcoin_5min.csv',
    'ethereum': './training_data/ethereum_5min.csv',
    'xau': './training_data/xau_5min.csv'
}
predictor.prepare_data(file_paths)

# Train model
train_losses, val_losses = predictor.train_model('bitcoin', epochs=100)

Loading Pre-trained Models

# Load a pre-trained model
predictor.load_model('bitcoin')

Prediction

# Make prediction using last 60 data points
last_60_data = predictor.data['bitcoin'].tail(60).to_dict('records')
prediction = predictor.predict_next_24h('bitcoin', last_60_data)

Real-Time Forecasting

from real_time_forecast import RealTimeForecaster

# Initialize real-time forecaster
forecaster = RealTimeForecaster()

# Run continuous forecasting (updates every minute)
forecaster.run_forecast()

Real-Time Forecasting System

The real-time forecasting system provides:

Features:

• Live Price Updates: Fetches current prices every minute
• 24-Hour Predictions: Generates 289 price points (current + 288 future)
• 5-Minute Intervals: Predictions at 5-minute intervals
• Multiple Assets: Bitcoin, Ethereum, and Gold (XAU)
• Automatic Saving: Saves forecasts to CSV files
• Continuous Operation: Runs indefinitely until stopped

Output:

• Console Display: Shows current prices and predictions
• CSV Files: Saves complete forecasts with timestamps
• Real-time Updates: Updates every 60 seconds

Sample Output:

🕐 Current time: 2024-01-15 14:30:00 UTC
Fetching current prices...
💰 bitcoin: $42,150.50
💰 ethereum: $2,450.75
💰 xau: $2,025.30

📊 Generating 24-hour forecasts...

🔮 Forecasting for BITCOIN
Current Price: $42,150.50
Current Time: 2024-01-15 14:30:00 UTC

First 10 predictions:
  14:35:00: $42,165.75
  14:40:00: $42,180.25
  14:45:00: $42,195.40
  ...

💾 Forecast saved to: ./forecasts/bitcoin_forecast_20240115_1430.csv
⏰ Next update in 60 seconds...

Model Saving

• Location: All trained models are saved in the ./models/ directory
• Naming: Models are named as best_model_{asset_name}.pth
• Automatic: Models are automatically saved during training when validation loss improves
• Loading: Models can be loaded using predictor.load_model(asset_name)

Files

• fetch_training_data.py: Fetches historical data from Pyth Network
• model_train.py: Main training script with LSTM model
• real_time_forecast.py: Real-time forecasting system
• test_data_fetch.py: Tests data format and quality
• test_real_time.py: Tests real-time forecasting functionality
• requirements.txt: Python dependencies
• training_data/: Directory containing fetched CSV data
• models/: Directory containing trained model files
• forecasts/: Directory containing real-time forecast outputs

Notes

• The model uses 5-minute intervals for training
• Predictions are made for the next 24 hours (288 intervals)
• Early stopping is implemented to prevent overfitting
• The system automatically handles data scaling and feature engineering
• Models are saved in PyTorch format (.pth) for easy loading and deployment
• Real-time system requires trained models and historical data to function
• Forecasts are saved with timestamps for easy tracking and analysis